Algoritmos de aprendizagem de máquina na modelagem da distribuição potencial de habitats de espécies arbóreas

Abstract

The aim of the present study was to evaluate three methods of machine learning (decision tree-J48, random forest and artificial neural networks) to model the potential habitat distribution of the ten most abundant tree species of the São Francisco river watershed. The presence/absence tree species data were from 77 fragments sampled with 2,234 plots. We used 12 categorical environmental variables from the Economic Ecological Zoning of Minas Gerais (ZEE/MG), as well as variables related to soil water balance (current and potential evapotranspiration, aridity and alpha index). The parameterization of the three algorithms was done with cv parameter algorithm of the WEKA software. The results showed the applied algorithms were statistically similar for 60% of the tree species. The random forest and multilayer perceptron algorithms were statistically similar considering the Eugenia dysenterica and superior to J48 algorithm. However, the random forest algorithm was superior to the other for the three species of Qualea genera. The conclusion is the random forest was the most robust model for the potential distribution habitat of tree species.